AuroraMiddleware/skia2
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Add a dynamic color attrib to hw tessellated stroking

Browse Source 
Only adds color to the hardware tessellator. The indirect tessellator
reorders draws with is log2 binning, so we can't have different
colors.

Bug: chromium:1172543
Bug: skia:10419
Change-Id: I2a3700cd4572e8222002bfb028af05c6ec447708
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/369976
Commit-Queue: Chris Dalton <csmartdalton@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This commit is contained in:
Chris Dalton 2021-02-18 15:22:54 -07:00 committed by Skia Commit-Bot
parent fac4efcabe
commit 1017a3558f
10 changed files with 165 additions and 60 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
bench/TessellateBench.cpp
View File

@ -231,7 +231,8 @@ private:
SkMatrix matrix = SkMatrix::Scale(fMatrixScale, fMatrixScale);
GrStrokeHardwareTessellator tessellator(ShaderFlags::kNone,
*fTarget->caps().shaderCaps());
tessellator.prepare(fTarget.get(), matrix, {fPath, fStrokeRec}, fPath.countVerbs());
tessellator.prepare(fTarget.get(), matrix, {fPath, fStrokeRec, SK_PMColor4fWHITE},
fPath.countVerbs());
}
}
@ -270,10 +271,10 @@ private:
for (int i = 0; i < loops; ++i) {
for (const SkPath& path : fPaths) {
GrStrokeIndirectTessellator tessellator(ShaderFlags::kNone, SkMatrix::I(),
{path, fStrokeRec}, path.countVerbs(),
fTarget->allocator());
tessellator.prepare(fTarget.get(), SkMatrix::I(), {path, fStrokeRec},
path.countVerbs());
{path, fStrokeRec, SK_PMColor4fWHITE},
path.countVerbs(), fTarget->allocator());
tessellator.prepare(fTarget.get(), SkMatrix::I(),
{path, fStrokeRec, SK_PMColor4fWHITE}, path.countVerbs());
}
fTarget->resetAllocator();
}

10
src/gpu/GrColor.h
View File

@ -92,13 +92,19 @@ static inline uint64_t SkPMColor4f_toFP16(const SkPMColor4f& color) {
*/
class GrVertexColor {
public:
explicit GrVertexColor(const SkPMColor4f& color, bool wideColor)
: fWideColor(wideColor) {
GrVertexColor() = default;
explicit GrVertexColor(const SkPMColor4f& color, bool wideColor) {
this->set(color, wideColor);
}
void set(const SkPMColor4f& color, bool wideColor) {
if (wideColor) {
memcpy(fColor, color.vec(), sizeof(fColor));
} else {
fColor[0] = color.toBytes_RGBA();
}
fWideColor = wideColor;
}
size_t size() const { return fWideColor ? 16 : 4; }

12
src/gpu/tessellate/GrStrokeHardwareTessellator.cpp
View File

@ -121,10 +121,17 @@ void GrStrokeHardwareTessellator::prepare(GrMeshDrawOp::Target* target, const Sk
auto tolerances = Tolerances::MakePreTransform(matrixScales.data(), stroke.getWidth());
this->updateTolerances(tolerances, stroke.getJoin());
}
fStroke = &stroke;
if (fShaderFlags & ShaderFlags::kDynamicStroke) {
fDynamicStroke.set(stroke);
}
fStroke = &stroke;
if (fShaderFlags & ShaderFlags::kDynamicColor) {
bool wideColor = fShaderFlags & ShaderFlags::kWideColor;
SkASSERT(wideColor || pathStroke.fColor.fitsInBytes());
fDynamicColor.set(pathStroke.fColor, wideColor);
}
fHasLastControlPoint = false;
SkDEBUGCODE(fHasCurrentPoint = false;)
SkPathVerb previousVerb = SkPathVerb::kClose;
@ -655,6 +662,9 @@ void GrStrokeHardwareTessellator::emitDynamicAttribs() {
if (fShaderFlags & ShaderFlags::kDynamicStroke) {
fPatchWriter.write(fDynamicStroke);
}
if (fShaderFlags & ShaderFlags::kDynamicColor) {
fPatchWriter.write(fDynamicColor);
}
}
bool GrStrokeHardwareTessellator::reservePatch() {

1
src/gpu/tessellate/GrStrokeHardwareTessellator.h
View File

@ -125,6 +125,7 @@ private:
// Stateful values for the dynamic state (if any) that will get written out with each patch.
GrStrokeTessellateShader::DynamicStroke fDynamicStroke;
GrVertexColor fDynamicColor;
friend class GrOp; // For ctor.

2
src/gpu/tessellate/GrStrokeIndirectTessellator.cpp
View File

@ -444,6 +444,8 @@ GrStrokeIndirectTessellator::GrStrokeIndirectTessellator(
const GrSTArenaList<PathStroke>& pathStrokeList, int totalCombinedVerbCnt,
SkArenaAlloc* alloc)
: GrStrokeTessellator(shaderFlags) {
// We can't combine colors because our log2 binning draws things out of order.
SkASSERT(!(fShaderFlags & ShaderFlags::kDynamicColor));
SkASSERT(!fTotalInstanceCount);
SkASSERT(!fResolveLevels);
SkASSERT(!fResolveLevelArrayCount);

54
src/gpu/tessellate/GrStrokeTessellateOp.cpp
View File

@ -20,13 +20,15 @@ GrStrokeTessellateOp::GrStrokeTessellateOp(GrAAType aaType, const SkMatrix& view
: GrDrawOp(ClassID())
, fAAType(aaType)
, fViewMatrix(viewMatrix)
, fColor(paint.getColor4f())
, fProcessors(std::move(paint))
, fPathStrokeList(path, stroke)
, fTotalCombinedVerbCnt(path.countVerbs()) {
, fPathStrokeList(path, stroke, paint.getColor4f())
, fTotalCombinedVerbCnt(path.countVerbs())
, fProcessors(std::move(paint)) {
if (SkPathPriv::ConicWeightCnt(path) != 0) {
fShaderFlags |= ShaderFlags::kHasConics;
}
if (!this->headColor().fitsInBytes()) {
fShaderFlags |= ShaderFlags::kWideColor;
}
SkRect devBounds = path.getBounds();
float inflationRadius = stroke.getInflationRadius();
devBounds.outset(inflationRadius, inflationRadius);
@ -62,21 +64,21 @@ GrProcessorSet::Analysis GrStrokeTessellateOp::finalize(const GrCaps& caps,
SkASSERT(fPathStrokeList.begin().fCurr->fNext == nullptr);
SkASSERT(fAAType != GrAAType::kCoverage || hasMixedSampledCoverage);
const GrProcessorSet::Analysis& analysis = fProcessors.finalize(
fColor, GrProcessorAnalysisCoverage::kNone, clip, &GrUserStencilSettings::kUnused,
hasMixedSampledCoverage, caps, clampType, &fColor);
this->headColor(), GrProcessorAnalysisCoverage::kNone, clip,
&GrUserStencilSettings::kUnused, hasMixedSampledCoverage, caps, clampType,
&this->headColor());
fNeedsStencil = !analysis.unaffectedByDstValue();
return analysis;
}
GrOp::CombineResult GrStrokeTessellateOp::onCombineIfPossible(GrOp* grOp, SkArenaAlloc* alloc,
const GrCaps&) {
const GrCaps& caps) {
using DynamicStroke = GrStrokeTessellateShader::DynamicStroke;
SkASSERT(grOp->classID() == this->classID());
auto* op = static_cast<GrStrokeTessellateOp*>(grOp);
if (fNeedsStencil ||
op->fNeedsStencil ||
fColor != op->fColor ||
fViewMatrix != op->fViewMatrix ||
fAAType != op->fAAType ||
fProcessors != op->fProcessors ||
@ -91,14 +93,29 @@ GrOp::CombineResult GrStrokeTessellateOp::onCombineIfPossible(GrOp* grOp, SkAren
// still decide to combine them.
combinedFlags |= ShaderFlags::kDynamicStroke;
}
if (combinedFlags & ShaderFlags::kDynamicStroke) {
// Don't actually enable dynamic stroke on ops that already have lots of verbs.
if (!this->shouldUseDynamicState(ShaderFlags::kDynamicStroke) ||
!op->shouldUseDynamicState(ShaderFlags::kDynamicStroke)) {
if (!(combinedFlags & ShaderFlags::kDynamicColor) && this->headColor() != op->headColor()) {
// The paths have different colors. We will need to enable dynamic color if we still decide
// to combine them.
combinedFlags |= ShaderFlags::kDynamicColor;
}
// Don't actually enable new dynamic state on ops that already have lots of verbs.
constexpr static GrTFlagsMask<ShaderFlags> kDynamicStatesMask(ShaderFlags::kDynamicStroke |
ShaderFlags::kDynamicColor);
ShaderFlags neededDynamicStates = combinedFlags & kDynamicStatesMask;
if (neededDynamicStates != ShaderFlags::kNone) {
if (!this->shouldUseDynamicStates(neededDynamicStates) ||
!op->shouldUseDynamicStates(neededDynamicStates)) {
return CombineResult::kCannotCombine;
}
}
// The indirect tessellator can't combine colors because its log2 binning draws things out of
// order. Only enable dynamic color if we have hardware tessellation.
if ((combinedFlags & ShaderFlags::kDynamicColor) && !this->canUseHardwareTessellation(caps)) {
return CombineResult::kCannotCombine;
}
fPathStrokeList.concat(std::move(op->fPathStrokeList), alloc);
fTotalCombinedVerbCnt += op->fTotalCombinedVerbCnt;
fShaderFlags = combinedFlags;
@ -136,14 +153,11 @@ void GrStrokeTessellateOp::prePrepareTessellator(GrPathShader::ProgramArgs&& arg
const GrCaps& caps = *args.fCaps;
SkArenaAlloc* arena = args.fArena;
// Only use hardware tessellation if the path has a somewhat large number of verbs. Otherwise we
// seem to be better off using indirect draws. Our back door for HW tessellation shaders isn't
// currently capable of passing varyings to the fragment shader either, so if the processors
// have varyings we need to use indirect draws.
// Only use hardware tessellation if we need dynamic color or if the path has a somewhat large
// number of verbs. Otherwise we seem to be better off using indirect draws.
GrStrokeTessellateShader::Mode shaderMode;
if (caps.shaderCaps()->tessellationSupport() &&
fTotalCombinedVerbCnt > 50 &&
!fProcessors.usesVaryingCoords()) {
if (this->canUseHardwareTessellation(caps) &&
((fShaderFlags & ShaderFlags::kDynamicColor) || fTotalCombinedVerbCnt > 50)) {
fTessellator = arena->make<GrStrokeHardwareTessellator>(fShaderFlags, *caps.shaderCaps());
shaderMode = GrStrokeTessellateShader::Mode::kTessellation;
} else {
@ -164,7 +178,7 @@ void GrStrokeTessellateOp::prePrepareTessellator(GrPathShader::ProgramArgs&& arg
}
auto* strokeTessellateShader = arena->make<GrStrokeTessellateShader>(
shaderMode, fShaderFlags, fViewMatrix, this->headStroke(), fColor);
shaderMode, fShaderFlags, fViewMatrix, this->headStroke(), this->headColor());
auto* fillPipeline = GrFillPathShader::MakeFillPassPipeline(args, fAAType, std::move(clip),
std::move(fProcessors));
auto fillStencil = &GrUserStencilSettings::kUnused;

33
src/gpu/tessellate/GrStrokeTessellateOp.h
View File

@ -24,9 +24,11 @@ public:
GrStrokeTessellator(ShaderFlags shaderFlags) : fShaderFlags(shaderFlags) {}
struct PathStroke {
PathStroke(const SkPath& path, const SkStrokeRec& stroke) : fPath(path), fStroke(stroke) {}
PathStroke(const SkPath& path, const SkStrokeRec& stroke, const SkPMColor4f& color)
: fPath(path), fStroke(stroke), fColor(color) {}
SkPath fPath;
SkStrokeRec fStroke;
SkPMColor4f fColor;
};
// Called before draw(). Prepares GPU buffers containing the geometry to tessellate.
@ -55,16 +57,25 @@ private:
DEFINE_OP_CLASS_ID
SkStrokeRec& headStroke() { return fPathStrokeList.head().fStroke; }
SkPMColor4f& headColor() { return fPathStrokeList.head().fColor; }
// Returns whether it is a good tradeoff to use the given dynamic state. Dynamic state improves
// batching, but if it isn't already enabled, it comes at the cost of having to write out more
// data with each patch or instance.
bool shouldUseDynamicState(ShaderFlags dynamicState) const {
// Use the dynamic state if either (1) the state is already enabled anyway, or (2) we don't
// Returns whether it is a good tradeoff to use the dynamic states flagged in the given
// bitfield. Dynamic states improve batching, but if they aren't already enabled, they come at
// the cost of having to write out more data with each patch or instance.
bool shouldUseDynamicStates(ShaderFlags neededDynamicStates) const {
// Use the dynamic states if either (1) they are all already enabled anyway, or (2) we don't
// have many verbs.
constexpr static int kMaxVerbsToEnableDynamicState = 50;
return (fShaderFlags & dynamicState) ||
(fTotalCombinedVerbCnt <= kMaxVerbsToEnableDynamicState);
bool anyStateDisabled = (bool)(~fShaderFlags & neededDynamicStates);
bool allStatesEnabled = !anyStateDisabled;
return allStatesEnabled || (fTotalCombinedVerbCnt <= kMaxVerbsToEnableDynamicState);
}
bool canUseHardwareTessellation(const GrCaps& caps) {
SkASSERT(!fStencilProgram && !fFillProgram); // Ensure we haven't std::moved fProcessors.
// Our back door for HW tessellation shaders isn't currently capable of passing varyings to
// the fragment shader, so if the processors have varyings we need to use indirect draws.
return caps.shaderCaps()->tessellationSupport() && !fProcessors.usesVaryingCoords();
}
const char* name() const override { return "GrStrokeTessellateOp"; }
@ -87,13 +98,11 @@ private:
const GrAAType fAAType;
const SkMatrix fViewMatrix;
SkPMColor4f fColor;
bool fNeedsStencil = false;
GrProcessorSet fProcessors;
ShaderFlags fShaderFlags = ShaderFlags::kNone;
GrSTArenaList<PathStroke> fPathStrokeList;
int fTotalCombinedVerbCnt = 0;
GrProcessorSet fProcessors;
bool fNeedsStencil = false;
GrStrokeTessellator* fTessellator = nullptr;
const GrProgramInfo* fStencilProgram = nullptr; // Only used if the stroke has transparency.

62
src/gpu/tessellate/GrStrokeTessellateShader.cpp
View File

@ -135,6 +135,9 @@ private:
// [NUM_RADIAL_SEGMENTS_PER_RADIAN, STROKE_RADIUS]
v->declareGlobal(GrShaderVar("vsStrokeArgs", kFloat2_GrSLType, TypeModifier::Out));
}
if (shader.hasDynamicColor()) {
v->declareGlobal(GrShaderVar("vsColor", kHalf4_GrSLType, TypeModifier::Out));
}
v->insertFunction(kAtan2Fn);
v->insertFunction(kCosineBetweenVectorsFn);
@ -380,12 +383,25 @@ private:
v->codeAppend(R"(
vsStrokeArgs = float2(NUM_RADIAL_SEGMENTS_PER_RADIAN, STROKE_RADIUS);)");
}
if (shader.hasDynamicColor()) {
v->codeAppend(R"(
vsColor = dynamicColorAttr;)");
}
// The fragment shader just outputs a uniform color.
const char* colorUniformName;
fColorUniform = uniHandler->addUniform(nullptr, kFragment_GrShaderFlag, kHalf4_GrSLType,
"color", &colorUniformName);
args.fFragBuilder->codeAppendf("%s = %s;", args.fOutputColor, colorUniformName);
if (!shader.hasDynamicColor()) {
// The fragment shader just outputs a uniform color.
const char* colorUniformName;
fColorUniform = uniHandler->addUniform(nullptr, kFragment_GrShaderFlag, kHalf4_GrSLType,
"color", &colorUniformName);
args.fFragBuilder->codeAppendf("%s = %s;", args.fOutputColor, colorUniformName);
} else {
// Color gets passed in from the tess evaluation shader.
SkString flatness(args.fShaderCaps->preferFlatInterpolation() ? "flat" : "");
args.fFragBuilder->declareGlobal(GrShaderVar(SkString("tesColor"), kHalf4_GrSLType,
TypeModifier::In, 0, SkString(),
flatness));
args.fFragBuilder->codeAppendf("%s = tesColor;", args.fOutputColor);
}
args.fFragBuilder->codeAppendf("%s = half4(1);", args.fOutputCoverage);
}
@ -423,7 +439,9 @@ private:
m.getScaleY());
}
pdman.set4fv(fColorUniform, 1, shader.fColor.vec());
if (!shader.hasDynamicColor()) {
pdman.set4fv(fColorUniform, 1, shader.fColor.vec());
}
}
GrGLSLUniformHandler::UniformHandle fTessArgsUniform;
@ -487,6 +505,10 @@ SkString GrStrokeTessellateShader::getTessControlShaderGLSL(
code.append(R"(
in vec2 vsStrokeArgs[];)");
}
if (this->hasDynamicColor()) {
code.append(R"(
in mediump vec4 vsColor[];)");
}
code.append(R"(
out vec4 tcsPts01[];
@ -500,6 +522,10 @@ SkString GrStrokeTessellateShader::getTessControlShaderGLSL(
code.append(R"(
patch out float tcsStrokeRadius;)");
}
if (this->hasDynamicColor()) {
code.append(R"(
patch out mediump vec4 tcsColor;)");
}
code.append(R"(
void main() {
@ -510,6 +536,10 @@ SkString GrStrokeTessellateShader::getTessControlShaderGLSL(
code.append(R"(
tcsStrokeRadius = vsStrokeArgs[0].y;)");
}
if (this->hasDynamicColor()) {
code.append(R"(
tcsColor = vsColor[0];)");
}
code.append(R"(
// Unpack the curve args from the vertex shader.
@ -831,6 +861,11 @@ SkString GrStrokeTessellateShader::getTessEvaluationShaderGLSL(
code.append(R"(
patch in float tcsStrokeRadius;)");
}
if (this->hasDynamicColor()) {
code.appendf(R"(
patch in mediump vec4 tcsColor;
%s out mediump vec4 tesColor;)", shaderCaps.preferFlatInterpolation() ? "flat" : "");
}
code.append(R"(
uniform vec4 sk_RTAdjust;)");
@ -932,7 +967,14 @@ SkString GrStrokeTessellateShader::getTessEvaluationShaderGLSL(
}
code.append(R"(
gl_Position = vec4(vertexPos * sk_RTAdjust.xz + sk_RTAdjust.yw, 0.0, 1.0);
gl_Position = vec4(vertexPos * sk_RTAdjust.xz + sk_RTAdjust.yw, 0.0, 1.0);)");
if (this->hasDynamicColor()) {
code.append(R"(
tesColor = tcsColor;)");
}
code.append(R"(
})");
return code;
@ -1242,9 +1284,11 @@ class GrStrokeTessellateShader::IndirectImpl : public GrGLSLGeometryProcessor {
void GrStrokeTessellateShader::getGLSLProcessorKey(const GrShaderCaps&,
GrProcessorKeyBuilder* b) const {
bool keyNeedsJoin = fMode == Mode::kIndirect && !(fShaderFlags & ShaderFlags::kDynamicStroke);
bool keyNeedsJoin = (fMode == Mode::kIndirect) && !(fShaderFlags & ShaderFlags::kDynamicStroke);
SkASSERT(fStroke.getJoin() >> 2 == 0);
uint32_t key = (uint32_t)fShaderFlags;
// Attribs get worked into the key automatically during GrPrimitiveProcessor::getAttributeKey().
// When color is in a uniform, it's always wide. kWideColor doesn't need to be considered here.
uint32_t key = (uint32_t)(fShaderFlags & ~ShaderFlags::kWideColor);
key = (key << 1) | (uint32_t)fMode;
key = (key << 2) | ((keyNeedsJoin) ? fStroke.getJoin() : 0);
key = (key << 1) | (uint32_t)fStroke.isHairlineStyle();

33
src/gpu/tessellate/GrStrokeTessellateShader.h
View File

@ -34,7 +34,9 @@ public:
enum class ShaderFlags {
kNone = 0,
kHasConics = 1 << 0,
kDynamicStroke = 1 << 1 // Each patch or instance has its own stroke width and join type.
kWideColor = 1 << 1,
kDynamicStroke = 1 << 2, // Each patch or instance has its own stroke width and join type.
kDynamicColor = 1 << 3, // Each patch or instance has its own color.
};
GR_DECL_BITFIELD_CLASS_OPS_FRIENDS(ShaderFlags);
@ -142,19 +144,23 @@ public:
// Size in bytes of a tessellation patch with the given shader flags.
static size_t PatchStride(ShaderFlags shaderFlags) {
size_t stride = sizeof(SkPoint) * 5;
if (shaderFlags & ShaderFlags::kDynamicStroke) {
stride += sizeof(DynamicStroke);
}
return stride;
return sizeof(SkPoint) * 5 + DynamicStateStride(shaderFlags);
}
// Size in bytes of an indirect draw instance with the given shader flags.
static size_t IndirectInstanceStride(ShaderFlags shaderFlags) {
size_t stride = sizeof(float) * 11;
return sizeof(float) * 11 + DynamicStateStride(shaderFlags);
}
// Combined size in bytes of the dynamic state attribs enabled in the given shader flags.
static size_t DynamicStateStride(ShaderFlags shaderFlags) {
size_t stride = 0;
if (shaderFlags & ShaderFlags::kDynamicStroke) {
stride += sizeof(DynamicStroke);
}
if (shaderFlags & ShaderFlags::kDynamicColor) {
stride += (shaderFlags & ShaderFlags::kWideColor) ? sizeof(float) * 4 : 4;
}
return stride;
}
@ -207,6 +213,13 @@ public:
fAttribs.emplace_back("dynamicStrokeAttr", kFloat2_GrVertexAttribType,
kFloat2_GrSLType);
}
if (fShaderFlags & ShaderFlags::kDynamicColor) {
fAttribs.emplace_back("dynamicColorAttr",
(fShaderFlags & ShaderFlags::kWideColor)
? kFloat4_GrVertexAttribType
: kUByte4_norm_GrVertexAttribType,
kHalf4_GrSLType);
}
if (fMode == Mode::kTessellation) {
this->setVertexAttributes(fAttribs.data(), fAttribs.count());
SkASSERT(this->vertexStride() == PatchStride(fShaderFlags));
@ -214,10 +227,12 @@ public:
this->setInstanceAttributes(fAttribs.data(), fAttribs.count());
SkASSERT(this->instanceStride() == IndirectInstanceStride(fShaderFlags));
}
SkASSERT(fAttribs.count() <= kMaxAttribCount);
}
bool hasConics() const { return fShaderFlags & ShaderFlags::kHasConics; }
bool hasDynamicStroke() const { return fShaderFlags & ShaderFlags::kDynamicStroke; }
bool hasDynamicColor() const { return fShaderFlags & ShaderFlags::kDynamicColor; }
private:
const char* name() const override { return "GrStrokeTessellateShader"; }
@ -237,7 +252,9 @@ private:
const ShaderFlags fShaderFlags;
const SkStrokeRec fStroke;
const SkPMColor4f fColor;
SkSTArray<4, Attribute> fAttribs;
constexpr static int kMaxAttribCount = 5;
SkSTArray<kMaxAttribCount, Attribute> fAttribs;
class TessellationImpl;
class IndirectImpl;

7
tests/StrokeIndirectTest.cpp
View File

@ -42,10 +42,11 @@ static void test_stroke(skiatest::Reporter* r, GrDirectContext* ctx, GrMockOpTar
float scale = ldexpf(rand.nextF() + 1, i);
auto matrix = SkMatrix::Scale(scale, scale);
GrStrokeIndirectTessellator tessellator(GrStrokeTessellateShader::ShaderFlags::kNone,
matrix, {path, stroke}, path.countVerbs(),
target->allocator());
matrix, {path, stroke, SK_PMColor4fWHITE},
path.countVerbs(), target->allocator());
tessellator.verifyResolveLevels(r, target, matrix, path, stroke);
tessellator.prepare(target, matrix, {path, stroke}, path.countVerbs());
tessellator.prepare(target, matrix, {path, stroke, SK_PMColor4fWHITE},
path.countVerbs());
tessellator.verifyBuffers(r, target, matrix, stroke);
}
}